Nvidia / 30 September 2026

DeepSeek and Huawei Challenge Nvidia’s CUDA Grip

DeepSeek has released open-source programming infrastructure developed with Huawei for the Chinese company’s Ascend AI chips, including computing and communication libraries. The pair also worked on a 128-chip Ascend 950 supernode. DeepSeek singled out TileLang, a high-level programming language, as a way to simplify development and called it “a simpler programming model” than Nvidia’s CUDA. The announcement turns China’s campaign to build domestic AI hardware into a more direct contest over the software developers use to make those chips productive. For Nvidia, CUDA is a source of customer familiarity and switching costs as well as a way to extract performance from GPUs. An alternative that works across large systems challenges that advantage at the programming layer.

DeepSeek said Huawei gave full support to the programming work, and the two companies optimized both calculation and communication on the supernode. A large model workload depends on how processors exchange data as well as how quickly one chip performs a task; the communications libraries therefore matter to the competitiveness of the whole cluster. Huawei unveiled its next generation of AI processors and supernode systems earlier in September and expects broad use for model training next year. The newly released tooling gives developers more of the software needed to translate model code into efficient execution on those systems. The practical contest will be measured in working applications and repeatable training and inference performance, not only the announced chip specifications.

TileLang sits above the low-level instructions that often make a hardware ecosystem difficult to adopt. DeepSeek argues that a high-level language can remain easy to program while reaching the processor’s full performance potential. Its statement called such a language a first priority for an independent GPU software ecosystem. This is a strategic distinction from simply porting one prominent model to Ascend: public compute and communication libraries allow other teams to adapt workloads, inspect the implementation and contribute optimizations. Developers with limited time are more likely to test an alternative accelerator when the software path is clear. Open sourcing also recruits outside engineering effort into a platform that Huawei would otherwise have to build alone.

China’s large cloud providers and AI labs need local sources of computing capacity as policy and trade restrictions complicate access to advanced Nvidia products. DeepSeek supplies influential model development expertise; Huawei supplies processors and systems. Their joint work joins these complementary assets around a public programming route rather than a private integration for one buyer. Nvidia’s existing CUDA ecosystem remains deep, with libraries, trained engineers and production tools accumulated over years. A credible domestic alternative can still change buyers’ negotiations before it matches every feature: a customer able to move some workloads has another option for capacity and pricing. The effect will vary by model, software compatibility and the cluster scale each provider can operate.

The 128-chip design is significant because it exposes communication and system-level behavior, areas in which attractive single-chip benchmarks do not settle a purchase decision. DeepSeek and Huawei have put specific software components into developers’ hands, creating a channel through which improvements can compound as applications appear. Nvidia faces a contest over the default development environment in a large market, even if its hardware continues to lead particular workloads. The story also travels beyond China: model teams elsewhere can evaluate portable high-level approaches when they allocate work among accelerator architectures. Greater software portability would make chip performance, system cost and supply terms more decisive in the next generation of purchases.